Optimizing Cervical Lesion Detection Using Deep Learning with Particle Swarm Optimization

Abstract

Cervical cancer ranks as the second most frequent malignancy in women worldwide. Cervigram is a type of cervical cancer screening tool that uses digital imaging technology. It captures images of the cervix to detect abnormalities or signs of cancerous growth. This method can assist in early detection and prompt treatment, potentially saving lives. Cervical is its reliance on human interpretation rather than automated analysis for detecting cervical abnormalities. In this study, we employed a metaheuristic approach PSO to enhance hierarchical feature learning in the Xception model for classifying cervigram images into cancerous and normal categories. By leveraging PSO, our approach optimizes the feature extraction process, while Xception’s deep learning capabilities facilitate intricate feature extraction, enabling precise differentiation between cancerous and normal cervigrams. We employ residual block with the Xception pre-trained model to facilitate faster convergence during training due to its ability to capture and refine complex hierarchical features. In this study, we utilized the benchmark Cervigram dataset, comprising 2779 Cancer and 1960 Normal images. The experiment demonstrates that our proposed model achieved an accuracy of 97.57%. Furthermore, our model outperforms existing studies and other pre-trained models.

Topics & Concepts